The right way to interpret this ruling is that a probable cause affidavit needs to offer evidence that a person has committed a crime, and model output is not evidence; it is more akin to an investigative tool.
Photographs, witness statements, direct observations, physical objects found at the scene are all evidence: They help answer questions such as "Was this person present at the scene?" or "Did this person take the actions they are alleged to have taken?"
A machine-learning model, even a very sophisticated one, is really at its core a probability machine. It might help you find two photographs that appear very similar, but it doesn't independently establish that the people in those photographs are, in fact, the same person. Likewise, an AI model might help you determine the similarity of speech patterns or writing styles, but those similarity scores don't actually tell you if an audio or writing sample came from the same person.
The person who submits the affidavit is vouching for the reasonableness of the facts they are presenting, and so any probabilistic evidence needs a human judgment call.
The wrong way to interpret this ruling is that AI can't ever be used as an investigative tool to establish probable cause. A detective certainly can use an algorithm that searches a facial-identification database, or a fingerprint database, to help identify potential suspects. But they then need to do something else to determine whether the model output actually matches the facts of the case. They can't just say, "The computer told me so."